Lightweight Object Tracking in Compressed Video Streams Demonstrated in Region-of-Interest Coding

Abstract

Video scalability is a recent video coding technology that allows content providers to offer multiple quality versions from a single encoded video file in order to target different kinds of end-user devices and networks. One form of scalability utilizes the region-of-interest concept, that is, the possibility to mark objects or zones within the video as more important than the surrounding area. The scalable video coder ensures that these regions-of-interest are received by an end-user device before the surrounding area and preferably in higher quality. In this paper, novel algorithms are presented making it possible to automatically track the marked objects in the regions of interest. Our methods detect the overall motion of a designated object by retrieving the motion vectors calculated during the motion estimation step of the video encoder. Using this knowledge, the region-of-interest is translated, thus following the objects within. Furthermore, the proposed algorithms allow adequate resizing of the region-of-interest. By using the available information from the video encoder, object tracking can be done in the compressed domain and is suitable for real-time and streaming applications. A time-complexity analysis is given for the algorithms proving the low complexity thereof and the usability for real-time applications. The proposed object tracking methods are generic and can be applied to any codec that calculates the motion vector field. In this paper, the algorithms are implemented within MPEG-4 fine-granularity scalability codec. Different tests on different video sequences are performed to evaluate the accuracy of the methods. Our novel algorithms achieve a precision up to 96.4%.